Segmentation Evaluation

Errors introduced during segmentation propagate through subsequent stages of iris pro-cessing, specifically normalization and feature extraction, confounding system performance.

Therefore, it would be advantageous to automatically detect such errors if they occurred. A relatively new avenue of research in iris literature is focused on evaluating the outcome of the segmentation block.

Lee [57] assumes that the gray level intensities within the pupil and iris regions are rela-tively uniform. Therefore, a favorable segmentation should maximize the uniformity within each region and maximize the contrast between different regions. This notion is captured by measuring intra-region homogeneity and inter-region heterogeneity; the former characterizes region uniformity while the latter characterizes differences between distinct regions. Lee employs eight distinct gradient operators to characterize the homogeneity and heterogeneity of the pupil and iris region. The estimated measures are combined through the application of PCA to generate a single segmentation prediction index. A similar approach was illus-trated by Zhou et al. [123] using local homogeneity measured from four regions along the iris boundary.

Zuo and Schmid [127] describe an approach which automatically evaluates the precision of iris segmentation. They argue that the gradient should be strong along the inner and outer iris boundaries for favorable segmentations. Therefore, a suitable descriptor for character-izing segmentation precision, should be a function of gradient strength along both contours while excluding regions which may be attenuated by occlusion. In practice, the gradient is derived from the polar representation of the iris along the angular variable. This measure is supplemented with a global constraint based on region intensity such that interior boundary

regions should be darker than exterior.

Kalka and Bartlow [49] utilize a combination of geometric and intensity features to de-cern wether an estimated segmentation is favorable or not. Given a segmentation, the pixels within the pupil boundary are classified as pupil or non-pupil pixels using a likelihood ra-tio test. This process is repeated while increasing the radius of the estimated pupil. The former metric captures information about over-segmentation while the latter can be used to characterize under segmentation. Both features are supplemented with a geometric feature that corresponds to the distance between the estimated pupil and iris center. In practice, the pupil and iris are not concentric but their centers are relatively within close proximately of each other. Therefore, the distance between both centers can be used as a simple feature to predict gross inaccuracies in the iris boundary. An overall evaluation result is gener-ated through the application of machine learning on the three aforementioned features (see Chapter 4for a thorough description and updates to this algorithm).

Iris Image quality

3.1 Introduction

In this chapter we introduce a comprehensive approach to assess quality from an iris bio-metric image. We identify a broad range of factors including defocus blur, off-angle, oc-clusion/specular reflection, lighting, and iris resolution. We then analyze their effect on traditional iris recognition systems. Publicly available iris data sets such as CASIA v3.0 [1], and ICE 1.0 [61] offer images with varying quality factors. The West Virginia University (WVU) data sets, also publicly available, have a broad range of quality factors present in images. West Virginia University (WVU) non-ideal [25], and West Virginia University off-angle (WVUOA) [2] are non-ideal iris image data sets. WVUOA is an off-off-angle data set collected in the NIR spectrum. This data set consists of only off-angle data which we utilize in the off-angle experiments in this paper. In order to get a data set with a broad range of factors we systematically collected an iris image data set of varying quality affected by defocus, occlusion/specular reflection, off-angle, lighting, and iris resolution. The intent

35

of this analysis is to evaluate how these factors influence recognition performance and to what extent. We are also aware of the availability of UBIRIS [80], the data set collected in the visible light spectrum. Traditional / commercial iris recognition systems operate in the NIR spectrum, typically 700 - 900nm. NIR imaging is used because it helps emphasize iris texture which can be difficult to image under visible light (i.e. texture is much richer for darkly pigmented eyes under NIR compared to visible). The main focus of this paper is to study and measure the impact of iris quality on traditional/commerical systems which utilize NIR imaging. Therefore, studying the quality factors in UBIRIS images is outside of the scope of this paper. Additionally, NIR is used the human eye is not as sensitive to NIR wavelengths as to visible light [27]. The pupil and eye remain open under NIR illumination.

However, a number of precautions has to be taken to ensure eye safety when dealing with NIR illumination. This is of particular concern in unconstrained environments when stronger NIR sources are required. The International standard for laser safety and equipment [21]

should be consulted when designing collection experiments. Exposure to wavelengths be-tween 700− 1050 should not exceed an irradiance of 10^{.0002(λ−700)}W m⁻² for captures lasting between 100− 1000 seconds.

Next we design procedures for estimation of defocus blur, motion blur, occlusion, specular reflection, lighting, off-angle, pixel-counts, and apply these to the images from data sets mentioned above. The individual factors are then “combined” using a fusion algorithm.

Although we evaluated several evidential reasoning fusion approaches [71], here we present the Dempster-Shafer criterion [73,74,96] which offers several advantages for combining the image quality scores. Fusing various iris quality attributes into a single “score” is somewhat controversial, yet it is practically important. Skeptics argue that the informative value of

a single fused iris quality score diminishes through fusion. For example, there is no doubt that if the goal of quality measurement is to improve the collection protocol, individual quality factors are more informative. On the other hand, vendors of biometric systems as well as the users desire to have a unified quality score as a measure of overall suitability for authenticating an individual. Further, recent “quality-enhanced” multi-modal biometric fusion algorithms [70,75] tend to use a single quality estimate for each biometric modality included in the fusion framework. Our work offers clear contributions in both of these contexts: (1) Scores that describe specific iris image quality attributes can be easily used for improving collection protocol. In fact, in this paper we compare the distribution of quality factors in each of the data sets listed above, thus offering an independent insight into the collection conditions. (2) We also demonstrate that the fused iris quality score clearly correlates with the matching score, which indicates that it can be used as an informal match confidence measure.